File size: 3,364 Bytes
e45f1b2 d87d9ba e45f1b2 771dd70 e45f1b2 6e691b3 4ee2327 e45f1b2 4ee2327 e45f1b2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 |
---
base_model: neuralmagic/Llama-2-7b-pruned50-retrained
inference: true
model_type: llama
pipeline_tag: text-generation
datasets:
- garage-bAInd/Open-Platypus
- Open-Orca/OpenOrca
- cognitivecomputations/dolphin
tags:
- sparse
- instruct
---
# Llama-2-7b-pruned50-retrained-instruct
This repo contains a [50% sparse Llama 2 7B](https://huggingface.co/neuralmagic/Llama-2-7b-pruned50-retrained) finetuned for instruction-following tasks using a blend of the Platypus + Open Orca + Dolphin datasets.
Official model weights from [Enabling High-Sparsity Foundational Llama Models with Efficient Pretraining and Deployment](https://arxiv.org/abs/2405.03594).
**Authors**: Neural Magic, Cerebras
## Usage
Below we share some code snippets on how to get quickly started with running the model.
### Sparse Transfer
By leveraging a pre-sparsified model's structure, you can efficiently fine-tune on new data, leading to reduced hyperparameter tuning, training times, and computational costs. Learn about this process [here](https://neuralmagic.github.io/docs-v2/get-started/transfer).
### Running the model
This model may be run with the transformers library. For accelerated inference with sparsity, deploy with [nm-vllm](https://github.com/neuralmagic/nm-vllm) or [deepsparse](https://github.com/neuralmagic/deepsparse).
```python
# pip install transformers accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Llama-2-7b-pruned50-retrained-instruct")
model = AutoModelForCausalLM.from_pretrained("Llama-2-7b-pruned50-retrained-instruct", device_map="auto")
input_text = "Write a recipe for banana bread:\n"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))
```
## Evaluation Benchmark Results
Model evaluation metrics and results.
| Benchmark | Metric | Llama-2-7b-instruct | Llama-2-7b-pruned50-retrained-instruct |
|------------------------------------------------|---------------|-------------|-------------------------------|
| [MMLU](https://arxiv.org/abs/2009.03300) | 5-shot | 48.60% | 45.10% |
| [HellaSwag](https://arxiv.org/abs/1905.07830) | 10-shot | 79.45% | 78.86% |
| [WinoGrande](https://arxiv.org/abs/1907.10641) | 5-shot | 75.69% | 72.61% |
| [ARC-c](https://arxiv.org/abs/1911.01547) | 25-shot | 53.92% | 50.77% |
| [TruthfulQA](https://arxiv.org/abs/2109.07958) | 0-shot | 43.63% | 44.40% |
| [GSM8K](https://arxiv.org/abs/2110.14168) | 5-shot | 15.92% | 16.38% |
## Model Training Details
This model was obtained by sparse-tranfer of the sparse foundational model [Llama-2-7b-pruned50-retrained](https://huggingface.co/neuralmagic/Llama-2-7b-pruned70-retrained) on a blend of [Open Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus), 10% [Open Orca](https://huggingface.co/datasets/Open-Orca/OpenOrca) and 10% [Dolphin](https://huggingface.co/datasets/cognitivecomputations/dolphin) datasets.
Training was perfomerd for 2 epochs.
## Help
For further support, and discussions on these models and AI in general, join [Neural Magic's Slack Community](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ) |